Inverse Scaling: When Bigger Isn't Better
read the original abstract
Work on scaling laws has found that large language models (LMs) show predictable improvements to overall loss with increased scale (model size, training data, and compute). Here, we present evidence for the claim that LMs may show inverse scaling, or worse task performance with increased scale, e.g., due to flaws in the training objective and data. We present empirical evidence of inverse scaling on 11 datasets collected by running a public contest, the Inverse Scaling Prize, with a substantial prize pool. Through analysis of the datasets, along with other examples found in the literature, we identify four potential causes of inverse scaling: (i) preference to repeat memorized sequences over following in-context instructions, (ii) imitation of undesirable patterns in the training data, (iii) tasks containing an easy distractor task which LMs could focus on, rather than the harder real task, and (iv) correct but misleading few-shot demonstrations of the task. We release the winning datasets at https://inversescaling.com/data to allow for further investigation of inverse scaling. Our tasks have helped drive the discovery of U-shaped and inverted-U scaling trends, where an initial trend reverses, suggesting that scaling trends are less reliable at predicting the behavior of larger-scale models than previously understood. Overall, our results suggest that there are tasks for which increased model scale alone may not lead to progress, and that more careful thought needs to go into the data and objectives for training language models.
This paper has not been read by Pith yet.
Forward citations
Cited by 15 Pith papers
-
AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents
AgentDojo introduces an extensible evaluation framework populated with realistic agent tasks and security test cases to measure prompt injection robustness in tool-using LLM agents.
-
LEVANTE-bench: Multi-Scale Comparison of VLMs to Children Using Cognitive Tasks (or, "Is Your VLM Smarter Than a 5th Grader?")
VLMs show partial alignment with children's performance on six cognitive tasks, with stronger models matching better at task and item levels but struggling on matrix reasoning and mental rotation.
-
The Variance Brain Foundation Models Forgot: Third-Order Statistics Predict Cognition Where Billion-Parameter Models Fail
Third-order co-skewness in fMRI is destroyed by BFM pretraining, causing poor cognition prediction; a co-skewness-preserving linear FC exceeds BFMs and raw FC.
-
GraphARC: A Comprehensive Benchmark for Graph-Based Abstract Reasoning
GraphARC is a scalable benchmark for few-shot graph transformation learning that exposes a comprehension-execution gap in language models on abstract reasoning tasks.
-
Evaluating Code Reasoning Abilities of Large Language Models Under Real-World Settings
A new dataset and nine-metric majority-vote procedure show that existing code-reasoning benchmarks are dominated by lower-complexity problems that do not reflect real-world code.
-
Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation
Chain-of-thought monitoring detects reward hacking in frontier reasoning models, but strong optimization against the monitor produces obfuscated misbehavior that remains hard to detect.
-
Trivium: Temporal Regret as a First-Class Objective for Causal-Memory Controllers
Trivium frames temporal regret as a first-class objective for causal-memory controllers and derives O(log E) bounds on miscalibration persistence under causal-probing assumptions, with preliminary validation on Causal...
-
Resonant Context Anchoring: Decoupling Attention Routing and Signal Gain at Inference Time
RCA is a training-free module that boosts input context signal strength in the residual stream of LLMs by orthogonal decoupling of attention routing from value magnitude.
-
History Anchors: How Prior Behavior Steers LLM Decisions Toward Unsafe Actions
A single consistency instruction with harmful prior actions causes aligned frontier LLMs to select unsafe options at 91-98% rates in high-stakes domains, with escalation and inverse scaling by model size.
-
The Granularity Gap: A Multi-Dimensional Longitudinal Audit of Sycophancy in Gemini Models
Continuous sycophancy scoring reveals that 27% of Gemini responses contain substantial social compliance missed by binary filters, with sycophancy correlating with hallucination and simple guardrails outperforming com...
-
EntroRouter: Learning Efficient Model Routing via Entropy Regulation
EntroRouter applies entropy regulation in a single-round routing framework to decouple reasoning from routing, retaining 98.3% of top expert accuracy at 48.25% lower compute cost.
-
From Verdict to Process: Agentic Reinforcement Learning for Multi-Stage Fact Verification
ProFact applies agentic RL with process-aware rewards to jointly optimize claim decomposition, evidence retrieval, and verdict prediction in multi-stage fact verification.
-
Absurd World: A Simple Yet Powerful Method to Absurdify the Real-world for Probing LLM Reasoning Capabilities
Absurd World automatically converts real-world problems into absurd yet logically coherent scenarios to test whether LLMs can reason without depending on familiar patterns.
-
Complexity Horizons of Compressed Models in Analog Circuit Analysis
Prerequisite graphs map compressed LLM performance boundaries in analog circuit analysis to allow selecting the smallest viable model for a given task complexity.
-
A Red Teaming Framework for Large Language Models: A Case Study on Faithfulness Evaluation
Introduces a multi-role red teaming framework using attacker and jury models that increases attack success rates by up to 7.9% on LLM faithfulness in question-answering tasks.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.